Windows Task Manager can now show you where your AI workloads are actually running. Until now, the Processes tab gave no per-process visibility into NPU (Neural Processing Unit) or GPU neural engine usage, so there was no standard way to verify whether a given AI app was truly running on the NPU. This is a change that affects both vendors shipping AI-powered apps and the teams supporting them internally.
🤖✍️ This article was co-written with AI — a Claude Code agent organized the primary sources and drafted the article, and Ebisuda reviewed and revised the content before publishing.
* New columns have been added to the Processes and Details tabs. Right-click the column header and add NPU, NPU engine, GPU, and GPU engine to see per-process execution targets. The Performance tab lets you track overall utilization over time. Support is limited to certain newer devices.
* Regular GPU compute is now distinguished from the "GPU neural engine." Recent silicon includes a dedicated matrix-multiplication accelerator inside the GPU itself, so you need to track separately whether a workload is landing on GPU 0 - Neural or on the NPU.
* Silent fallbacks become visible. Even when you've specified an execution target through ONNX Runtime or DirectML, a missing driver or a data-type mismatch (for example, loading an FP32 model onto an NPU that expects INT8) can cause the framework to silently fall back to the CPU or a regular GPU to avoid crashing. Typical symptoms are an NPU column stuck at 0%, an empty NPU engine field, or no neural indicator appearing in the GPU engine column.
* Dedicated and shared NPU memory are tracked separately. A rising trend in shared NPU memory usage is a sign that the dedicated pool has filled up and is starting to overflow, which leads to degraded performance. On integrated NPUs that have no dedicated memory pool, the dedicated-memory column will simply be empty.
* It also gives you a way to reason about power consumption. If the NPU engine column shows NPU - Compute or NPU - Neural, on-device AI processing is running on the NPU (the exact label varies by processor and driver version, but both mean the same thing). An app where only the CPU column is high while the NPU column sits idle has not been optimized. You can check the impact via the Power usage and Power usage trend columns.
From a help desk perspective, this gives you a way — without any additional tooling — to separate "the app itself is broken" from "it's not running on the execution target it was supposed to." From a development perspective, you can now confirm on the spot whether a quantization or binding fix actually had the intended effect.
* Task Manager features for visibility into AI workloads (Windows IT Pro Blog)
*
This article was written using "Ebi Workspace" (formerly claude-workspace), a paid plugin for Claude Code / Codex that provides the foundation for project management, context restoration, and skill execution.
Alongside it, I also run "Ebi Agent Chat Relay" (formerly CCDB; the repository name still reflects the old name as the rename is in progress), an open-source project that drives Claude Code / Codex from Discord and coordinates multiple sessions.
No posts
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.